Re: [PATCH] clocksource/arm_arch_timer: fix a lockdep warning

From: Qian Cai
Date: Mon Dec 03 2018 - 15:35:17 EST


On Mon, 2018-12-03 at 15:07 -0500, Waiman Long wrote:
> On 12/03/2018 02:33 PM, Qian Cai wrote:
> > Booting this Huawei TaiShan 2280 arm64 server generated this lockdep
> > warning.
> >
> > [ÂÂÂÂ0.000000]ÂÂlockdep_assert_cpus_held+0x50/0x60
> > [ÂÂÂÂ0.000000]ÂÂstatic_key_enable_cpuslocked+0x30/0xe8
> > [ÂÂÂÂ0.000000]ÂÂarch_timer_check_ool_workaround+0x128/0x2d0
> > [ÂÂÂÂ0.000000]ÂÂarch_timer_acpi_init+0x274/0x6ac
> > [ÂÂÂÂ0.000000]ÂÂacpi_table_parse+0x1ac/0x218
> > [ÂÂÂÂ0.000000]ÂÂ__acpi_probe_device_table+0x164/0x1ec
> > [ÂÂÂÂ0.000000]ÂÂtimer_probe+0x1bc/0x254
> > [ÂÂÂÂ0.000000]ÂÂtime_init+0x44/0x98
> > [ÂÂÂÂ0.000000]ÂÂstart_kernel+0x4ec/0x7d4
> >
> > This is due to the commit cb538267ea1e ("jump_label/lockdep: Assert we hold
> > the hotplug lock for _cpuslocked() operations"). Therefore, it will check
> > if it is really in the CPU hotplug path or not, and work around this
> > problem by using cpus_read_trylock(). The chance of not getting the read
> > lock is very small. If that happens, it will report a lockdep warning at
> > most.
> >
> > Signed-off-by: Qian Cai <cai@xxxxxx>
> > ---
> > Âdrivers/clocksource/arm_arch_timer.c | 9 +++++++++
> > Â1 file changed, 9 insertions(+)
> >
> > diff --git a/drivers/clocksource/arm_arch_timer.c
> > b/drivers/clocksource/arm_arch_timer.c
> > index 9a7d4dc..5c9acbd 100644
> > --- a/drivers/clocksource/arm_arch_timer.c
> > +++ b/drivers/clocksource/arm_arch_timer.c
> > @@ -497,11 +497,20 @@ void arch_timer_enable_workaround(const struct
> > arch_timer_erratum_workaround *wa
> > Â per_cpu(timer_unstable_counter_workaround, i) = wa;
> > Â }
> > Â
> > +#ifdef CONFIG_HOTPLUG_CPU
>
> If HOTPLUG_CPU isn't defined, all the cpus_lock() and related functions
> are just no-op. You don't need to use conditional compilation directive
> here.

Make sense.

>
> > + i = 0;
> > +
> > Â /*
> > Â Â* Use the locked version, as we're called from the CPU
> > Â Â* hotplug framework. Otherwise, we end-up in deadlock-land.
> > Â Â*/
>
> I think the main problem is the above comment may not be true anymore or
> is only occasionally true. We need to audit the code to find the root cause.

This was a commit introduced in Aug. 2017, 450f9689f294
(clocksource/arm_arch_timer: Use static_branch_enable_cpuslocked()) which
basically drop the cpus_read_lock(). May I ask what changes made you think the
above comment incorrect now?

>
> > + i = cpus_read_trylock();
> > Â static_branch_enable_cpuslocked(&arch_timer_read_ool_enabled);
> > + if (i)
> > + cpus_read_unlock();
>
> This is not the right way of fixing the lockdep splash.
>

I should had said it is a workaround. I am all-ears for a proper way to fix
this. When the above commitÂ450f9689f294 was merged, there was no cb538267ea1e
so no lockdep warning.